Know Today

September 23, 2026

AI agents Coding workflows Model economics Persistent context MCP integrations Visual prototyping Agent evaluation Automation skills

👀 Keep an eye on

Reported lower-cost frontier models make completed-task cost, not token price, the relevant routing metric.
Persistent folder-based context can turn scattered project knowledge into reusable operating memory for agents.
Scheduled skills plus parallel roles can connect intake, planning, implementation, testing, and visual inspection into one development loop.
Render-and-inspect workflows suggest coding agents can increasingly validate interactive and visual artifacts, not only produce code.
Small tools that extend established software may win adoption faster than AI-built replacements for mature workflows.

🧭 What changed

Reported model repricing: frontier capability is being presented at materially lower input and output prices, but task completion can still consume more tokens and tools.
Richer agent loops: planning, scheduled routines, parallel execution, and desktop-style preview are reported as converging in coding workflows.
Visual capability probes: interactive games, animations, emulation, and rendered scenes are becoming useful tests of end-to-end agent competence.

💭 Opinions worth testing

Opinion: Task-level cost is more useful than per-token pricing when choosing a model.

Opinion: Filesystem- and repository-centred workspaces are a stronger long-term base than isolated chat projects.

Opinion: Context quality and appropriately scoped permissions determine whether agents can perform useful operational work.

Opinion: Skills should be extracted from repeated successful work rather than designed speculatively.

Opinion: Thin orchestration around reliable skills is more valuable than a large fleet of loosely defined agents.

Opinion: Direct outcome testing deserves more weight than benchmark rank when evaluating models.

Opinion: Games and animation expose planning, iteration, and visual execution better than ordinary code snippets.

Opinion: Automated visual scores are useful signals, but human taste remains necessary for final experience quality.

Reflection: Instrument completed-task cost. Curate durable project context. Extract skills from proven runs. Build visual acceptance checks. Compare outputs against real user standards.

🔮 Predictions to track

Prediction: Teams with durable workspaces, reusable skills, and scheduled routines will automate a growing share of recurring work.

Prediction: Additive AI tools will see broader adoption than wholesale replacements of established SaaS workflows.

Prediction: More independent demos and evaluations will clarify whether newly reported model gains survive practical workloads.

Prediction: Model leadership will remain short-lived as successive flagship releases exceed current capability tiers.

Reflection: Track accepted-task quality. Re-run representative evaluations. Watch adoption inside existing workflows. Keep model routing configurable.

🛠️ Practical workflows

Create a project brief, an indexed knowledge folder, and reusable instructions that agents can load across sessions.
Run a recurring task manually until the output is dependable, then package the procedure as a parameterized skill with explicit completion metrics.
Use independent agents for implementation, testing, and design inspection, then join their work through clear quality gates.
Evaluate models on a representative suite covering coding, tool use, knowledge work, and one visual or interactive artifact.

🚀 Build something ambitious

Autonomous product-delivery operating system

Outcome: Convert customer signals into shipped product improvements with an evidence trail, lower coordination load, and measurable delivery economics.

Workflow: Ingest support, meeting, and product inputs; cluster opportunities; draft a plan; implement in isolated workspaces; run tests and visual checks; assemble a release brief; escalate only when authority, policy, or acceptance thresholds require a person.

Connected capabilities: Persistent project memory, email/calendar/chat connectors, MCP tools, planning agents, coding agents, browser or desktop preview, test runners, artifact generation, and task-level cost telemetry.

First meaningful milestone: Deliver one weekly feature candidate from incoming evidence to a tested change set with a visual acceptance report and per-accepted-task cost.

Compounding internal micro-tool factory

Outcome: Give teams a steady stream of narrowly useful interfaces that improve existing systems without forcing replacement.

Workflow: Detect repeated friction in operational data; create a lightweight artifact; collect inline feedback; convert accepted patterns into skills and templates; schedule recurring generation where the workflow stabilises.

Connected capabilities: Curated workspace context, connector-fed operational data, reusable skills, parameterized outputs, lightweight web artifacts, and feedback capture.

First meaningful milestone: Launch three no-login internal tools that each remove a recurring manual handoff and record usage plus feedback.

⚡ Fast builds to try

Model cost scoreboard: Run the same five real tasks through two models and log accepted output, latency, tokens, and total cost for an immediately usable routing baseline.
Project context map: Make a concise project brief plus an index of the most useful documents for a reusable agent starting point.
Visual QA loop: Ask an agent to build a small interactive page, capture its output, inspect the result against three criteria, and produce a clearer acceptance standard.
Feedback artifact: Generate a single-page concept demo with inline comments for a same-day stakeholder feedback channel.

⚠️ Caveats

Reported releases, pricing, availability, and benchmark comparisons need independent confirmation.
Visual demos and productivity claims are anecdotal and do not establish typical reliability, maintenance burden, or production readiness.
Connector-enabled workflows need explicit data governance, permission boundaries, retention rules, and escalation conditions before consequential actions.

🧠 The author's perspective

The author believes that today’s incoming information aligns with the view that increasingly autonomous software development and company operations will depend less on isolated coding prompts and more on a capable runtime with durable instructions, connected tools, decision data, and appropriately granted authority. Persistent workspaces supply the reusable context; skills and scheduled routines create repeatable operating loops; parallel agents divide planning, implementation, testing, and visual inspection; and connector-backed access makes agents useful beyond the repository. The reported move toward lower completed-task cost and closed-loop artifact validation strengthens the case for systems that learn from outcomes, retain reusable procedures, and ask people for help only at genuine authority, policy, or real-world interaction boundaries.